Local agent engine · Linux

An agent engine that shows its work.

Treebeard runs Qwen3.6-35B-A3B — a fast mixture-of-experts model — on hardware you own, with a tree-structured memory built for agents: fork conversations into waves, pick the best branch, and keep a receipt for every decision. One command installs the model, the runtime, and a 94/100-verified OpenAI-compatible server.

curl -fsSL https://raw.githubusercontent.com/newjordan/treebeard/main/install.sh | bash then  treebeard serve
94/100Agent Bench, reproduced on Intel + NVIDIA
262,144context tokens
12concurrent agent slots, one GPU
194 tok/s12-slot aggregate on Intel Arc Pro B70
3validated platforms: B70, GB10, portable CPU
Terminal demo: one-command Treebeard install, treebeard doctor showing 0.2.0-rc.1 pkg5 on Intel Arc Pro B70, treebeard serve starting the local API, and a first chat completion answering from the forest.
Why Treebeard

Better answers. Better economics. Better trust.

Every claim below is behind a preregistered gate with saved, checksum-pinned evidence — not a vibe.

◈

Better answers, measurably

Wave search forks a conversation into sibling branches and commits the self-consistent winner. Not best-of-N luck — structured search over your own model's candidates.

97.8% vs 53.6/63.6% single-pass, 30 evidence-bound receipts (PCBT-12)
0 regret deep-task quality gates — wave never loses to single decode (TQG-5)
⟁

Better economics of “better”

Best-of-N improves answers but costs N×. Treebeard's StateTree lets the branches share their prompt prefix, so search is nearly free — and many agents fit on one card.

2.9–8.4× cheaper than blind best-of-N, matched tasks (TQG)
+59–71% throughput on fragmented multi-branch workloads
+17–19% dense throughput at 8–12 agents (golden ABA)
✓

Better trust: receipts, not vibes

Every branch, winner-commit, and abort issues a pcbt.receipt.v1 receipt with attribution and evidence binding. Deterministic runs reproduce exactly across platforms.

Exact score + verdict-vector reproduction: Intel B70 ↔ NVIDIA GB10
Exact token parity across ragged/state-fused arms, drift ≤ 0.28%
How it works

A tree, not a treadmill

Agent work is branchy: plans, retries, parallel hypotheses. Treebeard's runtime treats conversation state as a tree instead of a single ever-growing tape.

01 / StateTree

Tree-structured KV memory

Sequence-ragged KV attention plus state-I/O fusion: forks share their prefix cells instead of copying them, and branches decode against the shared trunk. Shipped in the RC5–RC8 runtime line.

shared prefix→ fork()→ branch A branch B branch C
+58.9% ragged-only · +73.4% ship composition, fragmented shape

02 / Waves

Search over your own candidates

Fan a prompt out into a wave of forked siblings, decode each, score by verifier or self-consistency, commit the winner. The prefix is paid once; the search dividend is measured per gate.

prompt→ wave ×N→ score→ commit best
97.8% vs 53.6/63.6% single-pass (PCBT-12 product benchmark)

03 / PCBT

Proof-carrying branch transactions

The branch lifecycle is a transaction surface: client-driven attribution, evidence binding, atomic winner commit, abort and expiry — every transition receipted and counted in /metrics.

create→ bind evidence→ commit / abort→ receipt
pcbt.receipt.v1 · live on /transactions since RC8
Agent Bench

94/100, twice, on two different GPUs

69 deterministic agent scenarios, one server slot, 262,144-token context, temperature 0, seed 42 — reproduced exactly (score and verdict vector) on Intel Arc Pro B70 and NVIDIA GB10.

Agent Bench score
94
out of 100 · 130/138 points
Excellent
  • Scenarios69/69 completed
  • Outcomes63 pass · 4 partial · 2 fail
  • Request errors0
  • Cross-platformexact reproduction
A. Tool Selection6/6
B. Parameter Precision6/6
C. Multi-Step Chains8/8
D. Restraint & Refusal6/6
E. Error Recovery5/6
F. Localization6/6
G. Structured Reasoning6/6
H. Instruction Following8/10
I. Context & State17/20
J. Code Patterns6/6
K. Safety & Boundaries25/26
L. Toolset Scale8/8
M. Autonomous Planning5/6
N. Creative Composition6/6
O. Structured Output12/12

Machine-readable evidence: result.json · summary.json · nvidia-result.json · nvidia-summary.json · full results index

Selected measurements

Performance

MeasurementPlatformResult
Native pp4096NVIDIA GB102,026.9 tok/s
Native tg128NVIDIA GB1052.5 tok/s
Q8_0 direct 12-column speedupNVIDIA Blackwell33.0%
Q8_0 MoE down speedupNVIDIA Blackwell3.4%
12-slot aggregate servingIntel Arc Pro B70194.023 tok/s
Fragmented multi-branch compositionIntel Arc Pro B70 (StateTree)+73.4%
Dense golden 12-agent p50Intel Arc Pro B70 (StateTree)+19.16%
Installed-package chat smokeAMD Ryzen 9 5950X (portable CPU)9.30 tok/s
Releases

Now and next

current · pkg5

Treebeard 0.2.0

2026-07-18 · runtime 0.2.0-rc.1 · release notes
  • The StateTree runtime line goes public: sequence-ragged KV + state-I/O fusion, and the live PCBT transaction surface (pcbt.receipt.v1, /transactions).
  • Agent Bench 94/100 revalidated 2026-07-17 on Intel B70 and NVIDIA GB10 (exact score + verdict-vector reproduction); portable CPU build with perf parity.
  • PCBT-12 wave benchmark: 97.8% vs 53.6/63.6% single-pass; deep-task gates at 0 regret under the TQG-5 distinct-variant selector.
  • GB10 token generation at parity with the 0.1 line within host drift (three-point bisect, interleaved re-measurement).
previous · pkg4

Treebeard 0.1.1

2026-07-17 · runtime 0.1.0-rc.3 (launcher-only controls)
  • TREEBEARD_REASONING=off|bounded|unrestricted — thinking off by default (matches the 94/100 run); bounded budgets of 64 tokens (GPU) / 16 (CPU), overridable per request.
  • TREEBEARD_SPECULATION=off|ngram|mtp|hybrid — opt-in drafting with Qwen3.6's native MTP head; not yet claimed as a speedup.
Install

Running in three commands

About 26.7 GB, resumable, every byte SHA-256 verified. Needs roughly 32 GB of system or unified memory. Add --multimodal for the 0.9 GB vision projector.

curl -fsSL https://raw.githubusercontent.com/newjordan/treebeard/main/install.sh | bash

Downloads the model and the right runtime for your host — Intel SYCL, NVIDIA CUDA, or portable CPU.

treebeard doctor

Shows the platform selection and the exact launch command before anything starts.

treebeard serve

OpenAI-compatible API on http://127.0.0.1:8093/v1 — point your agent framework at it.